Back

Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures

J. BergstraDaniel YaminsDavid D. Cox

article2013ICML2,608 citations

Develops an automated Bayesian hyperparameter optimization framework that efficiently searches hundreds of architectural dimensions to match or surpass expert hand-tuning across standard vision benchmarks including CIFAR-10 and LFW.

Listen

Computer vision systems rely on many hyperparameters whose manual tuning often determines whether performance reaches state-of-the-art levels or remains near chance. Because these choices must be repeated for each new data set and are difficult to reproduce, it remains unclear whether reported gains reflect genuine algorithmic advances or simply better tuning.

The article set out to replace hand-tuning with an automated, reproducible search procedure that can optimize hundreds of interdependent hyperparameters across large families of feed-forward vision models.

The authors encoded a broad class of biologically inspired pipelines—covering filter generation, pooling, normalization, and classification—as an expression graph containing 238 hyperparameters. They compared two search strategies, random sampling and the Tree of Parzen Estimators algorithm, on the LFW face-verification, PubFig83 face-identification, and CIFAR-10 object-recognition tasks, running up to 2,000 evaluations per method.

TPE recovered or surpassed the best previously published configurations on all three data sets while using far fewer trials than random search. On LFW it reached 15.5 percent view-2 error versus 20.8 percent for random search; on PubFig83 it reached 13.5 percent versus 19.0 percent; and on CIFAR-10 it matched the accuracy of expert hand-tuning within roughly 800 trials. Random search never approached these levels within the allotted budget.

These results show that automated configuration can match or exceed skilled manual tuning at modest computational cost and, because every trial is logged, enables fair quantitative comparisons across modeling ideas. The approach therefore turns hyperparameter selection into a measurable component of model evaluation rather than an unquantified art.

The authors recommend encoding model families in searchable form and applying Bayesian optimization routines such as TPE or SMAC whenever new data sets or performance criteria arise. They note that further gains are likely from algorithms that capture hyperparameter interactions and from searches that explicitly trade accuracy against evaluation time or model size.

The main limitations are that TPE treats hyperparameters independently and that the validation sets used for search were not perfectly representative of the final test distributions, producing modest overfitting. Results are therefore most reliable when the same search protocol is repeated on held-out data or when larger computational budgets are available.

Massachusetts Institute of TechnologyThe Rowland Instituteoptimizationregularization parameterskernel parameterHyperparameter OptimizationFeed‑Forward Vision ArchitectureTree of Parzen Estimators
hyperopt/hyperopt-convnetBergstra et al (2013).pdf
Cover for Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures

Abstract

Many computer vision algorithms depend on configuration settings that are typically hand-tuned in the course of evaluating the algorithm for a particular data set. While such parameter tuning is often presented as being incidental to the algorithm, correctly setting these parameter choices is frequently critical to realizing a method's full potential. Compounding matters, these parameters often must be re-tuned when the algorithm is applied to a new problem domain, and the tuning process itself often depends on personal experience and intuition in ways that are hard to quantify or describe. Since the performance of a given technique depends on both the fundamental quality of the algorithm and the details of its tuning, it is sometimes difficult to know whether a given technique is genuinely better, or simply better tuned.

In this work, we propose a meta-modeling approach to support automated hyperparameter optimization, with the goal of providing practical tools that replace hand-tuning with a reproducible and unbiased optimization process. Our approach is to expose the underlying expression graph of how a performance metric (e.g. classification accuracy on validation examples) is computed from hyperparameters that govern not only how individual processing steps are applied, but even which processing steps are included. A hyperparameter optimization algorithm transforms this graph into a program for optimizing that performance metric. Our approach yields state of the art results on three disparate computer vision problems: a face-matching verification task (LFW), a face identification task (PubFig83) and an object recognition task (CIFAR-10), using a single broad class of feed-forward vision architectures.

Topics

optimizationregularization parameterskernel parameteroptimization problemssocial sciences

Keywords

Hyperparameter OptimizationFeed‑Forward Vision ArchitectureTree of Parzen EstimatorsRandom SearchLFWPubFig83CIFAR‑10Null Distribution Specification LanguageGaussian ProcessBayesian Optimization

Table of Contents

  • 1. Introduction
  • 2. Previous Work
  • 3. Automatic Hyperparameter Optimization
  • 4. Object Recognition Model Family
  • 5. Results
  • 5.1. TPE vs. Random Search: LFW and PubFig83
  • 5.2. Matching Hand-Tuning: CIFAR-10
  • 6. Discussion
  • 7. Acknowledgements
  • References

Knowls

  1. Knowl 1 — Four-Component Architecture for Graph-Based Hyperparameter Optimization

    model/method

    An automated hyperparameter optimization framework formalizes model search over complex, conditional configuration spaces via four interconnected components:

    1. Null Distribution Specification Language (GG): A domain-specific expression language that represents the hyperparameter search space as a directed acyclic expression graph (DAG). The nodes of GG specify prior probability distributions (such as normal, uniform, and log-uniform distributions) or deterministic transformations, and can be arbitrarily nested and referenced internally. It includes stochastic choice nodes that select among alternative sub-graphs, enabling the representation of conditional hyperparameters that are only evaluated when specific parent components are active.
    2. Loss Function: A black-box evaluation metric f:G→Rf: G \to \mathbb{R} that maps a sampled configuration from GG to a real scalar score (such as classification validation error). The loss function is computationally expensive and analytically intractable.
    3. Hyperparameter Optimization Algorithm (HOA): An iterative search procedure that takes the null prior graph GG and an experimental history H={(c1,y1),…,(ct,yt)}H = \{(c_1, y_1), \dots, (c_t, y_t)\} (where each cic_i is a configuration and yi=f(ci)y_i = f(c_i) is its evaluated loss) to propose subsequent configurations ct+1c_{t+1}. The HOA operates by replacing stochastic nodes within GG with adaptive conditional distributions estimated from HH.
    4. Database: A persistent storage layer recording the sequence of evaluated configuration-loss pairs HH. The database updates asynchronously as model evaluations finish, allowing the HOA to query the full experimental history to guide proposals.

  2. Knowl 2 — High-Dimensional Feed-Forward Computer Vision Architecture Search Space

    model/method

    The computer vision architecture search space defines a parameterized family of multi-stage feed-forward feature extractors coupled to a linear classifier, encompassing 238 hyperparameters across discrete architectural selections and continuous filtering/pooling options:

    • Pipeline Topology: Each pipeline consists of a sequence of 0, 1, or 2 inter-layers, followed by exactly 1 outer-layer, followed by column normalization and an ℓ2\ell_2-regularized linear Support Vector Machine (L2-SVM).
    • Inter-Layers: Each inter-layer performs filter-bank normalized cross-correlation (fbncc), local spatial pooling (lpool), and optional spatial sub-sampling (stride of 1 or 2). Filters are generated as random uniform filters, random projections of Zero-phase Component Analysis (ZCA) components, or ZCA-filtered image patches, parameterized by filter count K∈[16,256]K \in [16, 256], spatial filter size Sf∈[2,10]S_f \in [2, 10], and ZCA band-pass parameters.
    • Outer-Layers: The outer-layer applies fbncc where the filter count is dynamically determined so the total output feature dimension approaches but does not exceed 16,000. Feature pooling is selected via a conditional choice between:
      1. Local pooling (lpool) followed by divisor local normalization (lnorm).
      2. Difference of rectified histograms (dihist) aggregating positive and negative half-rectified filter responses, pooled either across a 2×22 \times 2 or 3×33 \times 3 spatial grid or via box filtering with sub-sampling strides of 1, 2, or 3 and window side-lengths from 2 to 8.
    • Classifier: A linear L2-SVM parameterized by an ℓ2\ell_2 regularization penalty CC and a variance threshold cutoff below which low-variance feature columns are removed.

  3. Knowl 3 — Modified Tree of Parzen Estimators with Aging and Adaptive Quantiles

    algorithm

    The Tree of Parzen Estimators (TPE) is a sequential model-based Bayesian optimization algorithm that models the conditional probability distributions ℓ(x)=P(x∣y<y∗)\ell(x) = P(x \mid y < y^*) and g(x)=P(x∣y≥y∗)g(x) = P(x \mid y \ge y^*) for each hyperparameter xx using Gaussian Mixture Models (GMMs). To handle high-dimensional vision architectures, TPE incorporates observation aging and an adaptive threshold quantile.

    Input: Null prior DAG specification GG, initial random trial count T0=50T_0 = 50, total budget Tmax⁡T_{\max}
    Output: Best configuration c∗c^* minimizing validation loss
    Initialize history H←∅H \leftarrow \emptyset
    for t=1t = 1 to T0T_0 do
        Sample configuration ct∼Gc_t \sim G
        Evaluate loss yt←LossFunction(ct)y_t \leftarrow \text{LossFunction}(c_t)
        H←H∪{(ct,yt)}H \leftarrow H \cup \{(c_t, y_t)\}
    for t=T0+1t = T_0 + 1 to Tmax⁡T_{\max} do
        Determine number of top trials kt←⌊t/4⌋k_t \leftarrow \lfloor \sqrt{t} / 4 \rfloor
        Sort HH in ascending order of loss: (c(1),y(1)),…,(c(t−1),y(t−1))(c_{(1)}, y_{(1)}), \dots, (c_{(t-1)}, y_{(t-1)})
        Set threshold y∗←y(kt)y^* \leftarrow y_{(k_t)}
        for each observation i∈{1,…,t−1}i \in \{1, \dots, t-1\} do
            if (t−i)≤25(t - i) \le 25 then
                wi←1.0w_i \leftarrow 1.0
            else
                wi←max⁡(0.0,1.0−(t−i−25)/(t−25))w_i \leftarrow \max(0.0, 1.0 - (t - i - 25) / (t - 25))
        Fit GMM ℓ(x)\ell(x) to hyperparameter values in {c(1),…,c(kt)}\{c_{(1)}, \dots, c_{(k_t)}\} using weights wiw_i
        Fit GMM g(x)g(x) to hyperparameter values in {c(kt+1),…,c(t−1)}\{c_{(k_t+1)}, \dots, c_{(t-1)}\} using weights wiw_i
        Sample candidate points from ℓ(x)\ell(x) and select ct=arg⁡max⁡cℓ(c)g(c)c_t = \arg\max_c \frac{\ell(c)}{g(c)}
        Evaluate loss yt←LossFunction(ct)y_t \leftarrow \text{LossFunction}(c_t)
        H←H∪{(ct,yt)}H \leftarrow H \cup \{(c_t, y_t)\}
    return c∗=arg⁡min⁡(c,y)∈Hyc^* = \arg\min_{(c, y) \in H} y

  4. Knowl 4 — Filter Bank Normalized Cross-Correlation Formulation

    equation

    The filter bank normalized cross-correlation operation (fbncc) applies a bank of KK spatial filters {fk}k=1K\{f_k\}_{k=1}^K to a multi-channel input feature map or image xx. For an input patch xˇij\check{x}_{ij} of spatial dimensions Sf×SfS_f \times S_f centered at row ii and column jj across all channels of xx, the patch is mean-adjusted according to: uˇij=xˇij−ϵmˇ\check{u}_{ij} = \check{x}_{ij} - \epsilon \check{m} where mˇ\check{m} is the empirical spatial-channel mean of xˇij\check{x}_{ij} and ϵ∈{0,1}\epsilon \in \{0, 1\} is a binary hyperparameter controlling whether patch mean subtraction is applied.

    The normalized filter response yijky_{ijk} for filter fkf_k is computed as: yijk=fk∗uˇijρmax⁡(∥uˇij∥22,β)+(1−ρ)(∥uˇij∥22+β)y_{ijk} = \frac{f_k * \check{u}_{ij}}{\sqrt{\rho \max(\|\check{u}_{ij}\|_2^2, \beta) + (1 - \rho)(\|\check{u}_{ij}\|_2^2 + \beta)}} where:

    • ∗* denotes spatial-channel correlation between filter fkf_k and patch uˇij\check{u}_{ij},
    • ∥uˇij∥2\|\check{u}_{ij}\|_2 is the Euclidean norm of patch uˇij\check{u}_{ij},
    • β>0\beta > 0 is a log-normally distributed variance floor hyperparameter preventing division by zero,
    • ρ∈{0,1}\rho \in \{0, 1\} is a binary hyperparameter selecting between a hard variance floor (when ρ=1\rho = 1) and a soft additive regularizer (when ρ=0\rho = 0).

  5. Knowl 5 — Spatial Pooling, Local Normalization, and Rectified Feature Representations

    equation

    In the feed-forward visual architecture family, post-filtering feature transformations are defined by three distinct mathematical operations:

    1. Local Spatial Pooling (lpool): For a patch surface xˇijk\check{x}_{ijk} of spatial dimension Sp×SpS_p \times S_p around spatial location (i,j)(i, j) in channel kk, the pooled response is normalized by its LpL_p norm: yijk=xi′j′k∥xˇi′j′k∥py_{ijk} = \frac{x_{i'j'k}}{\|\check{x}_{i'j'k}\|_p} where pp is a log-normally distributed continuous norm parameter, Sp∈[2,8]S_p \in [2, 8] is the spatial patch size, and (i′,j′)(i', j') denotes spatial subsampling with stride s∈{1,2}s \in \{1, 2\}.

    2. Divisive Local Normalization (lnorm): Normalizes the feature vector at spatial position (i,j)(i, j) across channels by its patch Euclidean norm if it exceeds a threshold: yij={xij∥xˇij∥2if ∥xˇij∥2>τxijotherwisey_{ij} = \begin{cases} \frac{x_{ij}}{\|\check{x}_{ij}\|_2} & \text{if } \|\check{x}_{ij}\|_2 > \tau \\ x_{ij} & \text{otherwise} \end{cases} where τ>0\tau > 0 is a log-normally distributed activation threshold hyperparameter.

    3. Difference Rectified Histogram Pooling (dihist): Computes spatial sums of positive and negative half-rectified responses over a local patch or grid cell: yijk=(∥max⁡(xˇijk−α,0)∥1∥max⁡(−xˇijk−α,0)∥1)y_{ijk} = \begin{pmatrix} \|\max(\check{x}_{ijk} - \alpha, 0)\|_1 \\ \|\max(-\check{x}_{ijk} - \alpha, 0)\|_1 \end{pmatrix} where α>0\alpha > 0 is a log-normally distributed rectification offset hyperparameter and ∥⋅∥1\|\cdot\|_1 denotes the ℓ1\ell_1 norm (spatial summation) over the pooling region.

  6. Knowl 6 — Empirical Optimization Performance of TPE on LFW and PubFig83

    empirical result

    When evaluated on face verification (Labeled Faces in the Wild, LFW) and face identification (PubFig83) using the 238-dimensional feed-forward model family:

    • Optimization Efficiency on LFW: On LFW View 1 validation, the Tree of Parzen Estimators (TPE) exceeds the best validation performance of a 2,000-trial random search within 200 trials and converges to an error rate of 16.2% within 1,000 trials (compared to 21.9% for random search).
    • Test Performance on LFW (View 2): TPE achieves a test accuracy of 84.5% (an error rate of 15.5±0.7%15.5 \pm 0.7\%) on View 2 test splits, outperforming a 2,000-trial random search (79.2%79.2\% accuracy / 20.8±0.8%20.8 \pm 0.8\% error) and matching high-throughput screening of 15,000 random models (15.9±0.7%15.9 \pm 0.7\% error).
    • Test Performance on PubFig83 (View 2): After screening over 1,200 model evaluations, TPE achieves an 83-way identification accuracy of 86.5% (13.50±0.7%13.50 \pm 0.7\% error rate), outperforming a 2,000-trial random search baseline (81.0%81.0\% accuracy / 19.0±0.8%19.0 \pm 0.8\% error) and exceeding the previous published state of the art (85.2%85.2\% accuracy / 14.78±0.45%14.78 \pm 0.45\% error).

  7. Knowl 7 — Test Error Rates on LFW and PubFig83 Benchmarks

    data/table

    The table compares test classification error rates (with 95% confidence intervals assuming Bernoulli-distributed errors) on the test sets ("View 2") for LFW (face verification) and PubFig83 (83-way face identification) across optimization approaches:

    Method (# configurations) LFW View 2 Error (%) PubFig83 View 2 Error (%)
    TPE-optimized (750) 15.5 ±\pm 0.7 13.50 ±\pm 0.7
    High-throughput (15K) 15.9 ±\pm 0.7 14.78 ±\pm 0.45
    Random search (2K) 20.8 ±\pm 0.8 19.0 ±\pm 0.8
    Chance 50.0 98.8

    The results demonstrate that TPE discovers architectures with lower test error rates in 750 trials than standard random search can find in 2,000 trials, and matches or exceeds the performance obtained through massive 15,000-trial random screening pipelines.

  8. Knowl 8 — Automated Hyperparameter Search Matches Expert Manual Tuning on CIFAR-10

    empirical result

    On the 10-way object recognition benchmark CIFAR-10 (32×3232 \times 32 color images, 50,000 training and 10,000 test images), automated hyperparameter optimization using the Tree of Parzen Estimators (TPE) over 800 trials achieved a test classification error of 21.2±0.8%21.2 \pm 0.8\%.

    This performance matches the 20.9±0.8%20.9 \pm 0.8\% test error achieved through domain expert hand-tuning of single-layer convolutional pipelines, while substantially outperforming a 2,000-trial random search baseline (23.4±0.8%23.4 \pm 0.8\% test error). The TPE search discovered an optimal single-layer configuration functionally similar to the hand-tuned architecture within roughly 24 hours of computation across 6 GPUs, with each candidate evaluation taking between 0 and 30 minutes.

  9. Knowl 9 — Classification Error Comparison on CIFAR-10 Benchmark

    data/table

    The table compares test classification error rates (with 95% confidence intervals assuming Bernoulli-distributed errors) on the CIFAR-10 object classification test set across different configuration search methods within the same architectural model class:

    Method (# configs) Test Error / Acc. (%)
    Hand-tuned 20.9 ±\pm 0.8
    TPE (800) 21.2 ±\pm 0.8
    Random (2K) 23.4 ±\pm 0.8
    Chance 90.0

    Note: The source document labels this metric column as "Test Acc. (%)", but the reported values (20.9%20.9\%, 21.2%21.2\%, 23.4%23.4\%, and chance level 90.0%90.0\%) represent test misclassification error percentages. The table shows that TPE matches expert hand-tuning within confidence margins after 800 evaluations, whereas 2,000 random search trials fail to reach competitive error levels.

  10. Knowl 10 — Limitations of Factorial Independence and Variable Computational Costs in TPE

    limitation

    The Tree of Parzen Estimators (TPE) algorithm exhibits two principal limitations when applied to large-scale vision architecture search:

    1. Factorial Independence Assumption: TPE models the joint distribution of hyperparameters conditionally independent of one another: P(c∣y)=∏iP(xi∣y)P(c \mid y) = \prod_i P(x_i \mid y) This assumption prevents TPE from modeling direct parameter interactions and correlations, despite optimal architectural settings often depending heavily on co-occurring configuration choices.
    2. Resource-Agnostic Proposals: TPE treats all loss evaluations as computationally uniform, even though evaluation wall times vary widely across configurations (ranging from near 0 to 30 minutes depending on filter size, layer depth, and pooling strides). It lacks mechanisms to optimize expected improvement per unit of computational time.

Coverage note — No substantial contributed material was omitted; the knowls cover the hyperparameter optimization framework, the visual model search space, the modified TPE algorithm, mathematical layer formulations, empirical evaluations on LFW, PubFig83, and CIFAR-10, and stated algorithmic limitations.

References

  1. 1.Bardenet, R. and Kéǵl, B. Surrogating the surrogate: accelerating Gaussian Process optimization with mixtures. In ICML, 2010.
  2. 2.Bergstra, J. Hyperopt: Distributed asynchronous hyperparameter optimization in Python. http://jaberg.github.com/hyperopt, 2013.
  3. 3.Bergstra, J. and Bengio, Y. Random search for hyperparameter optimization. Journal of Machine Learning Research, 13:281–305, 2012.
  4. 4.Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., Turian, J., and Bengio, Y. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010.
  5. 5.Bergstra, J., Bardenet, R., Bengio, Y., and Kéǵl, B. Algorithms for hyper-parameter optimization. In NIPS*24, pp. 2546–2554, 2011.
  6. 6.Bergstra, J., Pinto, N., and Cox, D. D. Machine learning for predictive auto-tuning with boosted regression trees. In INPAR, 2012.
  7. 7.Bergstra, J., Yamins, D., and Pinto, N. Hyperparameter optimization for convolutional vision architectures. https://github.com/jaberg/hyperopt-convnet, 2013.
  8. 8.Brochu, E. Interactive Bayesian Optimization: Learning Parameters for Graphics and Animation. PhD thesis, University of British Columbia, December 2010.
  9. 9.Coates, A. and Ng, A. Y. The importance of encoding versus training with sparse coding and vector quantization. In Proc. ICML-28, 2011.
  10. 10.DiCarlo, J. J., Zoccolan, D., and Rust, N. C. How does the brain solve visual object recognition? Neuron, 73:415–34, 2012 Feb 9 2012. ISSN 1097-4199.
  11. 11.Fan, R.-E., Chang, K.-W., Hsieh, C.-J., Wang, X.-R., , and Lin, C.-J. Liblinear: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874, 2008.
  12. 12.Fukushima, K. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–202, 1980.
  13. 13.Hinton, G. E., Osindero, S., and Teh, Y. A fast learning algorithm for deep belief nets. Neural Computation, 18:1527–1554, 2006.
  14. 14.Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller, E. Labeled faces in the wild: A database for studying face recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October 2007.
  15. 15.Hutter, F. Automated Configuration of Algorithms for Solving Hard Computational Problems. PhD thesis, University of British Columbia, 2009.
  16. 16.Hutter, F., Hoos, H., and Leyton-Brown, K. Sequential model-based optimization for general algorithm configuration. In LION-5, 2011. Extended version as UBC Tech report TR-2010-10.
  17. 17.Hyvarinen, A. and Oja, E. Independent component analysis: Algorithms and applications. Neural Networks, 13(4–5):411–430, 2000.
  18. 18.Jones, D.R. A taxonomy of global optimization methods based on response surfaces. Journal of Global Optimization, 21:345–383, 2001.
  19. 19.Krizhevsky, A. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009.
  20. 20.LeCun, Y., Boser, B., Denker, J. S., Henderson, D., Howard, R. E., Hubbard, W., and Jackel, L. D. Backpropagation applied to handwritten zip code recognition. Neural Computation, 1(4):541–551, 1989.
  21. 21.Lowe, D. G. Object recognition from local scale-invariant features. In Proceedings of the International Conference on Computer Vision 2 (ICCV), pp. 1150–1157, 1999. doi: 10.1109/ICCV.1999.790410.
  22. 22.Mockus, J., Tiesis, V., and Zilinskas, A. The application of Bayesian methods for seeking the extremum. In Dixon, L.C.W. and Szego, G.P. (eds.), Towards Global Optimization, volume 2, pp. 117–129. North Holland, New York, 1978.
  23. 23.Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., and Duchesnay, E. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12:2825–2830, 2011.
  24. 24.Pinto, N. and Cox, D. D. Beyond simple features: A large-scale feature search approach to unconstrained face recognition. In Proc. Face and Gesture Recognition, 2011.
  25. 25.Pinto, N., Doukhan, D., DiCarlo, J. J., and Cox, D. D. A high-throughput screening approach to discovering good forms of biologically inspired visual representation. PLoS Comput Biol, 5(11):e1000579, 11 2009.
  26. 26.Pinto, N., Stone, Z., Zickler, T., and Cox, D. D. Scaling-up Biologically-Inspired Computer Vision: A Case-Study on Facebook. In IEEE Computer Vision and Pattern Recognition, Workshop on Biologically Consistent Vision, 2011.
  27. 27.Rasmussen, C. E. and Williams, C. K. I. Gaussian Processes for Machine Learning. MIT Press, 2006.
  28. 28.Riesenhuber, M. and Poggio, T. Hierarchical models of object recognition in cortex. Nature Neuroscience, 2:1019–1025, 1999.
  29. 29.Snoek, J., Larochelle, H., and Adams, R. P. Practical bayesian optimization of machine learning algorithms. In Neural Information Processing Systems, 2012.

Citation

MLA
Bergstra, J., et al. “Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures”. Digital Access to Scholarship at Harvard (DASH) (Harvard University), 2013, pp. 115–23, http://nrs.harvard.edu/urn-3:HUL.InstRepos:12561000.
APA
Bergstra, J., Yamins, D., & Cox, D. (2013). Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures. Digital Access to Scholarship at Harvard (DASH) (Harvard University), 115–123. http://nrs.harvard.edu/urn-3:HUL.InstRepos:12561000
Chicago
Bergstra, J., D. Yamins, and D. Cox. 2013. “Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures”. Digital Access to Scholarship at Harvard (DASH) (Harvard University), 115–23. http://nrs.harvard.edu/urn-3:HUL.InstRepos:12561000.
Harvard
Bergstra, J., Yamins, D. and Cox, D. (2013) “Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures”, Digital Access to Scholarship at Harvard (DASH) (Harvard University), pp. 115–123. Available at: http://nrs.harvard.edu/urn-3:HUL.InstRepos:12561000.
Vancouver
1. Bergstra J, Yamins D, Cox D (2013) Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures. Digital Access to Scholarship at Harvard (DASH) (Harvard University) 115–123

BibTeX

@article{bergstra2013making,
  title = {Making a Science of Model Search: Hyperparameter Optimization in Hundreds of Dimensions for Vision Architectures},
  author = {Bergstra, James and Yamins, Daniel and Cox, David},
  year = {2013},
  journal = {Digital Access to Scholarship at Harvard (DASH) (Harvard University)},
  pages = {115-123},
  url = {http://nrs.harvard.edu/urn-3:HUL.InstRepos:12561000}
}
Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Authors