450 likes | 465 Views
This study explores efficient encoding of the world and its impact on visual and auditory processing. It also presents a model that learns hierarchical representations for early perceptual processing, leading to state-of-the-art object and face recognition.
E N D
Unsupervised learning of visual representations and their use in object & facerecognition Gary Cottrell Chris Kanan Honghao Shan Lingyun Zhang Matthew Tong Tim Marks
Chris Kanan Honghao Shan Collaborators
Lingyun Zhang Matt Tong Tim Marks Collaborators
Efficient Encoding of the world Sparse Principal Components Analysis: A model of unsupervised learning for early perceptual processing (Honghao Shan) The model embodies three constraints • Keep as much information as possible • While trying to equalize the neural responses • And minimizing the connections.
Persistent, small Midget? Trained on grayscale images Transient, large Temporal extent Parasol? Spatial extent Trained on video cubes Trained on color images Efficient Encoding of the world leads to magno- and parvo-cellular response properties… This suggests that these cell types exist because they are useful for efficiently encoding the temporal dynamics of the world.
Efficient Encoding of the world leads to gammatone filters as in auditory nerves: • Using exactly the same algorithm, applied to speech, environmental sounds, etc.:
Efficient Encoding of the world • A single unsupervised learning algorithm leads to • Model cells with properties similar to those found in the retina when applied to natural videos • Models cells with properties similar to those found in auditory nerve when applied to natural sounds • One small step towards a unified theory of temporal processing.
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • Recursive ICA (RICA 1.0 (Shan et al., 2008)): • Alternately compress and expand representation using PCA and ICA; • ICA was modified by a component-wise nonlinearity • Receptive fields expanded at each ICA layer
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • ICA was modified by a component-wise nonlinearity: • Think of ICA as a generative model: The pixels are the sum of many independent random variables: Gaussian. • Hence ICA prefers its inputs to be Gaussian-distributed. • We apply an inverse cumulative Gaussian to the absolute value of the ICA components to “gaussianize” them.
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • Strong responses, either positive or negative, are mapped to the positive tail of the Gaussian; weak ones, to the negative tail; ambiguous ones to the center.
SPCA SPCA Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • RICA 2.0: • Replace PCA by SPCA
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • RICA 2.0 Results: Multiple layer system with • Center-surround receptive fields at the first layer • Simple edge filters at the second (ICA) layer • Spatial pooling of orientations at the third (SPCA) layer: • V2-like response properties at the fourth (ICA) layer
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • V2-like response properties at the fourth (ICA) layer • These maps show strengths of connections to layer 1 ICA filters. Warm and cold colors are strong +/- connections, gray is weak connections, orientation corresponds to layer 1 orientation. • The left-most column displays two model neurons that show uniform orientation preference to layer-1 ICA features. • The middle column displays model neurons that have non-uniform/varying orientation preference to layer-1 ICA features. • The right column displays two model neurons that have location preference, but no orientation preference, to layer-1 ICA features. The left two columns are consistent with Anzen, Peng, & Van Essen 2007. The right hand column is a prediction
Unsupervised Learning of Hierarchical Representations (RICA 2.0; cf. Shan et al., NIPS 19) • Dimensionality Reduction & Expansion might be a general strategy of information processing in the brain. • The first step removes noise and reduces complexity, the second step captures the statistical structure. • We showed that retinal ganglion cells and V1 complex cells may be derived from the same learning algorithm, applied to pixels in one case, and V1 simple cell outputs in the second. • This highly simplified model of early vision is the first one that learns the RFs of all early visual layers, using a consistent theory - the efficient coding theory. • We believe it could serve as a basis for more sophisticated models of early vision. • An obvious next step is to train and thus make predictions about higher layers.
Nice, but is it useful? We showed in Shan & Cottrell (CVPR 2008) that we could achieve state-of-the-art face recognition with the non-linear ICA features and a simple softmax output. We showed in Kanan & Cottrell (CVPR 2010) that we could achieve state-of-the-art face and object recognition with a system that used an ICA-based salience map, simulated fixations, non-linear ICA features, and a kernel-density memory. Here I briefly describe the latter.
Our attention is automatically drawn to interesting regions in images. Our salience algorithm is automatically drawn to interesting regions in images. These are useful locations for discriminating one object (face, butterfly) from another. One reason why this might be a good idea…
Main Idea Training Phase (learning object appearances): Use the salience map to decide where to look. (We use the ICA salience map) Memorize these samplesof the image, with labels (Bob, Carol, Ted, or Alice) (We store the (compressed) ICA feature values)
Main Idea Testing Phase (recognizing objects we have learned): Now, given a new face, use the salience map to decide where to look. Compare new image samples to stored ones - the closest ones in memory get to vote for their label.
Stored memories of Bob Stored memories of Alice New fragments Result: 7 votes for Alice, only 3 for Bob. It’s Alice! 19
Voting • The voting process is based on Bayesian updating (with Naïve Bayes). • The size of the vote depends on the distance from the stored sample, using kernel density estimation. • Hence NIMBLE: NIM with Bayesian Likelihood Estimation.
Overview of the system The ICA features do double-duty: They are combined to make the salience map - which is used to decide where to look They are stored to represent the object at that location
Image Global Features Global Classifier Decision NIMBLE vs. Computer Vision • Compare this to (most, not all!) computer vision systems: • One pass over the image, and global features.
Belief After 1 Fixation Belief After 10 Fixations
Robust Vision • Human vision works in multiple environments - our basic features (neurons!) don’t change from one problem to the next. • We tune our parameters so that the system works well on Bird and Butterfly datasets - and then apply the system unchanged to faces, flowers, and objects • This is very different from standard computer vision systems, that are (usually) tuned to a particular domain
Cal Tech 101: 101 Different Categories AR dataset: 120 Different People with different lighting, expression, and accessories
~7 fixations required to achieve at least 90% of maximum performance
So, we created a simple cognitive model that uses simulated fixations to recognize things. • But it isn’t that complicated. • How does it compare to approaches in computer vision?
Caveats: As of mid-2010. Only comparing to single feature type approaches (no “Multiple Kernel Learning” (MKL) approaches). Still superior to MKL with very few training examples per category.
1 5 15 30 NUMBER OF TRAINING EXAMPLES
1 2 3 6 8 NUMBER OF TRAINING EXAMPLES
Again, best for single feature-type systems and for 1 training instance better than all systems
More neurally and behaviorally relevant gaze control and fixation integration. • People don’t randomly sample images. • A foveated retina • Comparison with human eye movement data during recognition/classification of faces, objects, etc.
A biologically-inspired, fixation-based approach can work well for image classification. Fixation-based models can achieve, and even exceed, some of the best models in computer vision. …Especially when you don’t have a lot of training images.
Software and Paper Available at www.chriskanan.com For more details email: ckanan@ucsd.edu This work was supported by the NSF (grant #SBE-0542013) to the Temporal Dynamics of Learning Center.
Sparse Principal Components Analysis • We minimize: • Subject to the following constraint:
The SPCA model as a neural net… It is AT that is mostly 0…
Results • suggesting the 1/f power spectrum of images is where this is coming from…
Results • The role of : • Recall this reduces the number of connections…
Results • The role of : higher means fewer connections, which alters the contrast sensitivity function (CSF). • Matches recent data on malnourished kids and their CSF’s: lower sensitivity at low spatial frequencies, but slightly better at high than normal controls…
NIMBLE represents its beliefs using probability distributions Simple nearest neighbor density estimation to estimate: P(fixationt | Category = k) Fixations are combined over fixations/time using Bayesian updating