Comparison of machine learning architectures for neural decoding of self-location

---

In my master dissertation project at the University of Nottingham, a set of tools has been developed to decode the position, head orientation and velocity of an animal from neurophysiological signals recorded from raw place cells. The program can use behaviour cloning, transfer learning or CNN-BiLSTM architecture for machine learning.

My thesis can be read HERE

Data

Raw neural recordings from CA1 in rodents using tetrode recordings from Markus Frey et. al.(2021) experiment: https://figshare.com/articles/dataset/Tetrode_recordings_from_CA1/14909766?backTo=/collections/DeepInsight_-_Data_sharing/5486703

Features

• read raw signal from a .nwb or a .h5 files
• convert raw .nwb file into a raw .h5 format file in the required data system
• preprocess files with optional bandpass, line noise, detrend, rolling window mean filters and downsampling
• write preprocessed files into a .h5 file
• use single CPU, paralell computing(on multi-core CPU) or use GPU for preprocessing, based on the fastest calculated processing mode
• learn decoding from neural code by:
  • Behaviour cloning
  • Transfer learning (Efficientnet B0 with imagenet weights)
  • CNN-BiLSTM architecture
• use preprocessed data from ram or file for training or preprocess on-the-flight
• save best weights
• make predictions with the trained network
• evaulate model (save results)

Example project

In the example.ipynb file you can find an example project.

How to run

To run the project, follow these steps:

1. Clone or download this repository to your local machine.
2. Install the required packages listed in the requirements.txt file.
3. Open the example.ipynb file and edit the variables and settings as needed.
4. Run the segments and wait for the results.

Variables and Settings

The following variables and settings are used in the project:

Variable	Description	Possible Values	Value Type
base_path	Gives the source folder where the sources were saved	Any valid folder path	String
is_test	If it's True no output data will be saved and make a minimum runtime to check functions	True or False	Boolean
is_save	If it's True all results will be saved (eg.: plots, models, etc.)	True or False	Boolean
source_file_path	Gives the source path (base_path + 'source file name')	Any valid file path	String
original_srate	Sampling rate of the raw signal in Hz	30000 (based on your signal)	Integer
led_srate	Sampling rate of the raw LED signal in Hz	1000 (based on your signal)	Integer
new_srate	Downscaled target sampling rate in Hz	10000 (based on your signal)	Integer
window_size	Size of one window in milliseconds	2000 (based on your signal)	Integer
chunk_size	Size of one chunk in milliseconds	30000 (based on your signal)	Integer
is_save_images	If it is True the plots will be saved	True or False	Boolean
image_save_dir	The name of the target folder where the plots will be saved	Any valid folder path	String
start_point	It gives the start point of the plotted signal parts in ms	Any positive integer	Integer
end_point	It gives the end point of the plotted signal parts in ms	Any positive integer greater than start_point	Integer
tasks	A dictionary containing the required preprocessing steps and their settings. The possible keys are 'Cheby_band', 'narrow_filt', 'down_samp', 'detrend', 'roll_mean', 'normalize', and 'feature_extract'.	If you want to switch off a preprocessing step, write False to the parameters section. See default settings below.	False or a list
'Cheby_band'	Chebyshev bandpass filter parameters: [lowcut, highcut, order]	Normally [2,250,5] or any valid filter parameters within the range of the signal frequency spectrum. The order must be a positive integer. The lowcut and highcut must be positive floats. The lowcut must be lower than the highcut.	List of floats and integer
'narrow_filt'	Notch filter parameters: [notch frequency, quality factor]	Normally [50,20] or any valid filter parameters within the range of the signal frequency spectrum. The notch frequency must be a positive float. The quality factor must be a positive float greater than zero.	List of floats
'down_samp'	Downsample parameter: [new sampling rate]	[new_srate] or any valid sampling rate lower than the original sampling rate. The new sampling rate must be a positive integer.	List of integer
'detrend'	Detrend parameter: True or False. If it is True, a linear detrending will be applied to the signal.	True or False	Boolean
'roll_mean'	Rolling mean parameter: [windows factor]. If it is a positive integer greater than zero, a rolling mean with a window size of windows factor times the window size will be applied to the signal. If it is zero, no rolling mean will be applied.	Normally [20]. Any positive integer or zero.	Integer

Future Improvements

I'm constantly working to improve this project and make it even better! Here are some areas I'm planning to focus on in the near future:

• More visualisation features: We're exploring ways to add additional plotting features to make this project even more useful and powerful.
• Better documentation: I working on a more detailed documentation with more example projects and description to use all the features.
• Bug fixes and performance improvements: No software is perfect, and I'm always on the lookout for ways to improve the performance and stability of this project.
• Creating a Python package: I am currently working on turning the project into an installable python package, including the body functions.
• Fit to principles: Along with the previous point, the next update will rewrite the code to better fit the Interface Segregation and the Dependency Inversion Principles.

I'm always looking for ways to make our project better, and I welcome your feedback and suggestions.

License

This project is licensed under the MIT License - see the LICENSE file for details.