Keras CNN model accuracy not improving and decreasing over epoch?

Question

Newbie to machine learning here. I'm currently working on a diagnostic machine learning framework using 3D-CNNs on fMRI imaging. My dataset consists of 636 images right now, and I'm trying to distinguish between control and affected (binary classification). However, when I tried to train my model, after every epoch, my accuracy remains at 48.13%, no matter what I do. Additionally, over the epoch, the accuracy decreases from 56% to 48.13%. So far, I have tried:

• changing my loss functions (poisson, categorical cross entropy, binary cross entropy, sparse categorical cross entropy, mean squared error, mean absolute error, hinge, hinge squared)
• changing my optimizer (I've tried Adam and SGD)
• changing the number of layers
• using weight regularization
• changing from ReLU to leaky ReLU (I thought perhaps that could help if this was a case of overfitting)

Nothing has worked so far.

Any tips? Here's my code:

#importing important packages
import tensorflow as tf
import os
import keras
from keras.models import Sequential
from keras.layers import Dense, Flatten, Conv3D, MaxPooling3D, Dropout, BatchNormalization, LeakyReLU
import numpy as np
from keras.regularizers import l2
from sklearn.utils import compute_class_weight
from keras.optimizers import SGD

BATCH_SIZE = 64
input_shape=(64, 64, 40, 20)

# Create the model
model = Sequential()

model.add(Conv3D(64, kernel_size=(3,3,3), activation='relu', input_shape=input_shape, kernel_regularizer=l2(0.005), bias_regularizer=l2(0.005), data_format = 'channels_first', padding='same'))
model.add(MaxPooling3D(pool_size=(2, 2, 2)))
model.add(Conv3D(64, kernel_size=(3,3,3), activation='relu', input_shape=input_shape, kernel_regularizer=l2(0.005), bias_regularizer=l2(0.005), data_format = 'channels_first', padding='same'))
model.add(MaxPooling3D(pool_size=(2, 2, 2)))
model.add(BatchNormalization(center=True, scale=True))

model.add(Conv3D(64, kernel_size=(3,3,3), activation='relu', input_shape=input_shape, kernel_regularizer=l2(0.005), bias_regularizer=l2(0.005), data_format = 'channels_first', padding='same'))
model.add(MaxPooling3D(pool_size=(2, 2, 2)))
model.add(Conv3D(64, kernel_size=(3,3,3), activation='relu', input_shape=input_shape, kernel_regularizer=l2(0.005), bias_regularizer=l2(0.005), data_format = 'channels_first', padding='same'))
model.add(MaxPooling3D(pool_size=(2, 2, 2)))
model.add(BatchNormalization(center=True, scale=True))

model.add(Flatten())
model.add(BatchNormalization(center=True, scale=True))
model.add(Dense(128, activation='relu', kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01)))
model.add(Dropout(0.5))
model.add(Dense(128, activation='sigmoid', kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01)))
model.add(Dense(1, activation='softmax', kernel_regularizer=l2(0.01), bias_regularizer=l2(0.01)))
 
# Compile the model
model.compile(optimizer = keras.optimizers.sgd(lr=0.000001), loss='poisson', metrics=['accuracy', tf.keras.metrics.Precision(), tf.keras.metrics.Recall()])

# Model Testing 
history = model.fit(X_train, y_train, batch_size=BATCH_SIZE, epochs=50, verbose=1, shuffle=True)

Answer 1

The main issue is that you are using softmax activation with 1 neuron. Change it to sigmoid with binary_crossentropy as a loss function.

At the same time, bear in mind that you are using Poisson loss function, which is suitable for regression problems not classification ones. Ensure that you detect the exact scenario that your are trying to solve.

Answer 2

Softmax with one neuron makes the model illogical and only use one of the sigmoid activation functions or Softmax in the last layer