Pattern recognition of deep and superficial layers of the macaque brain using large-scale local field potentials

A 2019 conference paper tests whether machine learning can tell deep from superficial cortical layers using local field potentials from macaque monkeys. The recordings span all six layers.

“Pattern recognition of deep and superficial layers of the macaque brain using large-scale local field potentials,” by Omar Costilla-Reyes, André M. Bastos, and Earl K. Miller, was presented as a poster at the 2019 Conference on Cognitive Computational Neuroscience (CCN), held in Berlin on September 13–16, 2019. It comes from Dr. Costilla Reyes’s Picower postdoctoral fellowship in MIT Brain and Cognitive Sciences, where he applied machine learning to neural signals, learning, memory, and cortical dynamics.

The cerebral cortex is organized in six anatomical layers. The superficial layers (2 and 3) and the deep layers (5 and 6) are thought to express different neuronal rhythms and to carry different roles in feedforward and feedback communication. Local field potentials (LFPs) capture population-level electrical activity over time from electrodes in the tissue, so a probe with contacts spanning the cortical depth records one signal per depth. Telling deep layers from superficial ones in such recordings is a central problem, and a pattern-recognition method that infers layer from the LFP signal itself would make large recording sets easier to analyze consistently.

The authors recorded LFPs with multilaminar probes across all six cortical layers simultaneously in macaque monkeys, in the prefrontal cortex (PFC) and visual area V4, during a delayed match-to-sample task, in which the animal holds a sample stimulus in memory and later judges whether a test stimulus matches it. An ensemble machine-learning approach, which combines several models, identified the layers with good accuracy in both areas. Telling the layers apart from the recorded signal alone opens the way to analyzing layer-specific neural activity across large recording sets without an anatomical check after every experiment.

A companion paper at the same conference, led by Bastos with Costilla-Reyes and Miller, “Automatic methods for cortex-wide layer identification of electrophysiological signals reveals a cortical motif for the expression of neuronal rhythms,” applied automatic layer identification across cortical areas. In April 2019 Dr. Costilla Reyes also gave a tutorial at MIT on neural decoding of spike trains and local field potentials with machine learning in Python.