Deep Learning: Branching into brains

  1. Adam Shai
  2. Matthew Evan Larkum  Is a corresponding author
  1. Stanford University, United States
  2. Humboldt University, Germany

Abstract

What can artificial intelligence learn from neuroscience, and vice versa?

Main text

Deep learning is a subfield of machine learning that focuses on training artificial systems to find useful representations of inputs. Recent advances in deep learning have propelled the once arcane field of artificial neural networks into mainstream technology (LeCun et al., 2015). Deep neural networks now regularly outperform humans on difficult problems like face recognition and games such as Go (He et al., 2015; Silver et al., 2017). Traditional neuroscientists have also taken an interest in deep learning because it seemed initially that there were telling analogies between deep networks and the human brain. Nevertheless, there is a growing impression that the field might be approaching a new ‘wall’ and that deep networks and the brain are intrinsically different.

Chief among these differences is the widely held belief that backpropagation, the learning algorithm at the heart of modern artificial neural networks, is biologically implausible. This issue is so central to current thinking about the relationship between artificial and real brains that it has its own name: the credit assignment problem. The error in the output of a neural network (that is, the difference between the output and the 'correct' answer) can be reported or 'backpropagated' to any connection in the network, no matter where it is, to teach the network how to refine the output. But for a biological brain, neurons only receive information from the neurons they are connected to, making credit assignment a real problem. How does the brain blindly adjust the strength of the connections between neurons that are far removed from the output of the network? In the absence of a solution, we may be forced to conclude that deep learning and brains are incompatible after all.

Now, in eLife, Jordan Guerguiev, Timothy Lillicrap and Blake Richards propose a biologically inspired solution to the credit assignment problem (Guerguiev et al., 2017). Central to their model is the structure of the pyramidal neuron, which is the most prevalent cell type in the cortex (the outer layer of the brain). Pyramidal neurons have been a source of aesthetic pleasure and interesting research questions for neuroscientists for decades. Each neuron is shaped like a tree with a trunk reaching up and dividing into branches near the surface of the brain as if extending toward a source of energy or information. Can it be that, while most cells of the body have relatively simple shapes, evolution has seen to it that cortical neurons are so intricately shaped as to be apparently impractical?

Guerguiev et al. – who are based at the University of Toronto, the Canadian Institute for Advanced Research, and DeepMind – report that this impractical shape has an advantage: the long branched structure means that error signals at one end of the neuron and sensory input at the other end are kept separate from each other. These sources of information can then be brought together at the right moment in order to find the best solution to a problem.

As Guerguiev et al. note, many facts about real neurons and the structure of the cortex turn out to be just right to find optimal solutions to problems. For instance, the bottoms of cortical neurons are located just where they need to be to receive signals about sensory input, while the tops of these neurons are well placed to receive feedback error signals (Cauller, 1995; Larkum, 2013). The key to this design principle seems to be to keep these distinct information streams largely independent. At the same time, ion channels under the control of a host of other nearby neurons process and gate the transfer of information within the neuron.

Taking inspiration from these facts Guerguiev et al. implement a deep network with units that have different compartments, just like real neurons, that can separate sensory input from feedback error signals. These units have all the information they need to know in order to nudge the network toward the desired output. Guerguiev et al. prove formally that this approach is mathematically sound. Moreover, their new, biologically plausible deep network is able to perform well on a task to identify handwritten numbers, and does so by creating what are referred to as hierarchical representations. This phenomenon refers to the increasingly complex nature of the responses of the network's layers, commonly found in more traditional deep learning models, and in the sensory cortices of biological brains.

Doubtless, there will be more twists and turns to this story as more biological details are incorporated into the model. For instance the brain also faces a time-based credit assignment problem (Friedrich et al., 2011; Gütig, 2016). Guerguiev et al. admit that this network does not outperform non-biologically derived deep networks – yet. Nevertheless, the model they present paves the way for future work that links biological networks to machine learning. The hope is that this can be a two-way process, in which insights from the brain can be used to improve artificial intelligence, and insights from artificial intelligence can be used to reveal how the brain operates.

References

  1. Conference
    1. He K
    2. Zhang X
    3. Ren S
    4. Sun J
    (2015)
    Delving deep into rectifiers: Surpassing human-level performance on imagenet classification
    Proceedings of the IEEE International Conference on Computer Vision. pp. 1026–1034.

Article and author information

Author details

  1. Adam Shai

    Adam Shai is in the Department of Biology, Stanford University, Stanford, United States

    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1833-3906
  2. Matthew Evan Larkum

    Matthew Evan Larkum is at the Neurocure Cluster of Excellence, Humboldt University of Berlin, Germany

    For correspondence
    matthew.larkum@hu-berlin.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9799-2656

Publication history

  1. Version of Record published: December 5, 2017 (version 1)

Copyright

© 2017, Shai et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 5,982
    Page views
  • 673
    Downloads
  • 3
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Adam Shai
  2. Matthew Evan Larkum
(2017)
Deep Learning: Branching into brains
eLife 6:e33066.
https://doi.org/10.7554/eLife.33066

Further reading

    1. Computational and Systems Biology
    2. Epidemiology and Global Health
    Bitya Raphael-Mizrahi et al.
    Research Article

    The endocannabinoid system consists mainly of 2-arachidonoylglycerol and anandamide, as well as cannabinoid receptor type 1 (CB1) and type 2 (CB2). Based on previous studies, we hypothesized that a circulating peptide previously identified as Osteogenic Growth Peptide (OGP) maintains a bone-protective CB2 tone. We tested OGP activity in mouse models and cells, and in human osteoblasts. We show that the OGP effects on osteoblast proliferation, osteoclastogenesis, and macrophage inflammation in vitro, as well as rescue of ovariectomy-induced bone loss and prevention of ear edema in vivo are all abrogated by genetic or pharmacological ablation of CB2. We also demonstrate that OGP binds at CB2 and may act as both an agonist and positive allosteric modulator in the presence of other lipophilic agonists. In premenopausal women, OGP circulating levels significantly decline with age. In adult mice, exogenous administration of OGP completely prevented age-related bone loss. Our findings suggest that OGP attenuates age-related bone loss by maintaining a skeletal CB2 tone. Importantly, they also indicate the occurrence of an endogenous peptide that signals via CB2 receptor in health and disease.

    1. Cancer Biology
    2. Computational and Systems Biology
    Iurii Petrov, Andrey Alexeyenko
    Research Article

    Late advances in genome sequencing expanded the space of known cancer driver genes several-fold. However, most of this surge was based on computational analysis of somatic mutation frequencies and/or their impact on the protein function. On the contrary, experimental research necessarily accounted for functional context of mutations interacting with other genes and conferring cancer phenotypes. Eventually, just such results become 'hard currency' of cancer biology. The new method, NEAdriver employs knowledge accumulated thus far in the form of global interaction network and functionally annotated pathways in order to recover known and predict novel driver genes. The driver discovery was individualized by accounting for mutations' co-occurrence in each tumour genome - as an alternative to summarizing information over the whole cancer patient cohorts. For each somatic genome change, probabilistic estimates from two lanes of network analysis were combined into joint likelihoods of being a driver. Thus, ability to detect previously unnoticed candidate driver events emerged from combining individual genomic context with network perspective. The procedure was applied to ten largest cancer cohorts followed by evaluating error rates against previous cancer gene sets. The discovered driver combinations were shown to be informative on cancer outcome. This revealed driver genes with individually sparse mutation patterns that would not be detectable by other computational methods and related to cancer biology domains poorly covered by previous analyses. In particular, recurrent mutations of collagen, laminin, and integrin genes were observed in the adenocarcinoma and glioblastoma cancers. Considering constellation patterns of candidate drivers in individual cancer genomes opens a novel avenue for personalized cancer medicine.