Papers
Alex Bäcker
Research report · Caltech Center for Neuromorphic Systems Engineering
The figures of the original report are not reproduced here; their captions are kept in place.
Abstract. I am interested in understanding how the brain performs cognitive functions that still prove conceptually challenging to implement on computers, and in applying biological insights to develop novel algorithms. I am also working on developing bioinformatics and neuroinformatic analysis tools to help interpret the wealth of new data coming in through advances in genomics, high-throughput gene expression assays and multi-electrode neural recordings.
My work in the past few years has focused mostly in the olfactory system as a model system for pattern recognition and learning. I have found that early olfactory circuits appear to employ plasticity, gain control and temporal coding across synchronizing neuronal assemblies to solve the odor recognition problem across multiple concentrations.
Rationale for CNSE strategic plan. My research has focused on object recognition, learning and olfaction. Object recognition is a task fundamental to many industrial applications which humans continue to perform substantially better than machines. Man-made machines are still notoriously bad at learning — they are still usually custom-built for every application. Olfactory sensors have undergone substantial development in the past few years, but algorithms for exploiting them to extract identity and spatial information about odor sources in a manner robust to concentration changes are still lacking. My research may contribute to bridge these gaps.
Research summary. Object recognition requires both specificity, to ensure that stimuli with distinct behavioral relevance are distinguished, and invariance, to ensure that different instances of the same stimulus are recognized as the same under varied conditions (intensity, pitch, position, …). Psychophysical studies show that an odor can be perceived as identical over significant ranges of concentrations. Whether concentration invariance results, at least in part, from low-level neural phenomena rather than cognitive grouping is so far unknown.
I explored, firstly, the contribution of projection neurons (PNs) in the antennal lobe of the locust, the analog of the vertebrate olfactory bulb, to the recognition of odor identity across concentrations; and secondly, what role spike timing, neuronal identity, and synchronization among neuronal assemblies play in the encoding and decoding of odor information by downstream neurons.
I designed and built a novel computerized odor delivery system (Figs. 1, 2), which I used to show the following:
Figure 6. Computerized odor delivery system and electrophysiological rig
Figure 7. Software to control odor delivery system.
The locust can recognize odors, and shows innate olfactory preferences.
PNs solve the task of encoding both odorant concentration and odorant identity, independently of concentration, in three ways. First, by multiplexing information in different response dimensions using a code that involves neuronal identity, spike timing (on a timescale slower than previously believed) and synchronization across a neuronal assembly. Second, via a novel phenomenon of experience-dependent plasticity that contributes to PNs’ invariance to concentration and sensitizes PNs after exposure to an odor at high concentration, contrary to the adaptation exhibited by receptors. Third, a phenomenon of gain control, whereby excitatory and inhibitory responses balance out massive changes in receptor activity as a function of odorant concentration, maintains the output of PNs within a small dynamic range.
Figure 8. Experience sensitizes the locust olfactory system to previously perceived odors and makes responses more similar across concentrations, potentially facilitating recognition.
Figure 9. Visual priming facilitates recognition: a robust stimulus (bottom) facilitates subsequent recognition of a degraded version of the same stimulus (top).
A further mechanism of gain control contributing to keep the activity of early olfactory circuits relatively constant across the wide dynamic range of odorant concentrations in the air is the physical chemistry of odorant reception confers the olfactory system invariance to odorant volatility, a physical property that has hitherto been believed to play a fundamental role in an odorant’s effectiveness.
Response patterns sometimes exhibit stable representations over large composition ranges and then abrupt transitions as a function of concentration and mixture composition, suggesting the difference between "same" and "different" odors may be delineated by sharp boundaries in odor space.
Finally, how is the distributed code for odors in PN assemblies decoded? I showed that although synchronization among PN assemblies does not augment stimulus information in PN temporal responses, it is necessary for the read-out of odor information by downstream neurons.
In related work, I showed that the variability of PN responses is correlated across neurons (with Mike Wehr), how to produce non-cyclic Winnerless Competition (WLC) and a learning rule that causes random networks to self-organize into WLC (with Valentin Zhigulin), introduced an exact hypothesis test for binomial distributions, improvements to sliding-window cross-correlation (with Stijn Cassenaer) and to the K-nearest neighbor classification algorithm, a combinatorial analysis of the connectivity between the locust antennal lobes and mushroom bodies (with Gilles Laurent) and demonstrated heterogeneity along the length of the locust antenna (with Mike Wehr).
In an exploration of memory in visual search, I demonstrated, with Eva Peral, shortcomings in what was previously believed to constitute evidence for a lack of memory in visual search.
Finally, in collaboration with Eva Peral, I developed an eyetracker that combines image processing with learning neural networks to produce a remote video-based eyetracker that requires no special illumination or head gear.
Figure 10. A pupil and eye corners found by a remote video eye-tracker employing a neural network downstream of an image processing algorithm
Refereed Publications
Bäcker, A. & Laurent, G.: Priming in early olfactory circuits. Submitted.
MacLeod, K., Bäcker, A. & Laurent, G.: Who reads the temporal information in synchronized spike trains? Nature 395, 693 - 698 (1998).
Conference Proceedings
Bäcker, A: Gain control in early olfactory circuits. Soc for Neurosci Abstr , 2002.
Bäcker, A. & Laurent, G.: Concentration priming in projection neuron responses in the locust olfactory system. Soc for Neurosci Abstr , 2000.
Bäcker, A. & Peral, E.: Does visual search have memory?: A mathematical framework suggests current experimental paradigm may not be conclusive. Vision Research Conference, 1999.
Bäcker, A., MacLeod, K., Wehr, M. & Laurent, G.: Disruption of neuronal synchronization impairs reliable reconstruction of odors by beta lobe cells but not by projection neurons of the antennal lobe of the locust. Soc for Neurosci Abstr , 1998.
Bäcker, A., Wehr, M. & Laurent, G.: Analysis of the variability of temporal response patterns of simultaneously-recorded projection neurons in the olfactory system of the locust. Soc for Neurosci Abstr , 1997.
Bäcker, A & Peral, E.: A remote video-based eye tracker. Caltech CNS/EE 148 Technical Report, 1997.
Patents
Bäcker, A.: Computational Method To Find Interactions Between Genes Using Multiple Genomes. U.S. Provisional Patent.
Bäcker, A.: Automatic Methods to find Associations between Expressions. U.S. Provisional Patent.
Bäcker, A.: Method To Automatically Associate A Web Page Or Document With An Expression And Summarize Documents. U.S. Provisional Patent.
Bäcker, A.: Method To Direct Users To Relevant Portion Of A Web Document. U.S. Provisional Patent.
Bäcker, A.: User-Specific Adaptive Method To Sort Search Results By Relevance For The World Wide Web And Other Interconnected Data Structures. U.S. Provisional Patent.
Bäcker, A, Peral, E. & Munich, M.: A Remote Video-Based Eye Tracker. U.S. Provisional Patent.