The cell of a perceptron was based on an oversimplified conception of a neuron. At the time, neural plasticity, timing factors in relation to activation, neurochemical pathways, and energy transit complexities in axons were unknown. The mapping of pulse transmission to basic algebra seemed unrealistic, so timing was ignored. Plasticity, timing, and regional brain signaling are still neglected in most artificial network chapters in computer science textbooks, which is why multi-layer perceptrons are generally considered a type of artificial neural network even though there is almost nothing neural about them.
Recurrent networks were an attempt move in the direction of neuroscience, although not cognitive. Their goal was still within the confines of machine learning and not the wider conception of intelligence. LSTMs continued the machine learning disciplines along a path toward neural networks found in brains by trying to simulate the features of long and short term memory in the cells of the artificial network. The justification was that RNNs could be configured to learn short term temporal patterns or long term temporal patterns, but would not learn from data that required a balance between both.
It is possible that some combination of LSTMs, CNNs, attention based networks, RBMs, and other machine learning components will produce artificial cognition, but cognition involves more than learning a function of several variables and time. One of the known phenomena in cognitive science is that neural networks will adjust to a local brain injury so that the function lost can be partly reacquired, but only partly. This indicates that structure is plastic but not like an array in memory that can be filled with a process that can be re-run. The tissue of the brain is tuned for even the long term learned behavior.
This is in opposition to the idea of general intelligence in the brain and more indicative of a more general intelligence in biology with a specific and higher speed extension of that intelligence in animal brains. Cognition is not specific to humans, as primates and other mammals can produce cognitive responses. They cannot perform extended logic, develop powerful analogies, or communicate tool production procedures to family and friends.
It was believed that long term memory was obtained through structures in the organelles of neurons. Neural imaging has yet to produce the detail necessary to make a determination. Neural plasticity is surely involved in long term learning. Short term memory has been proposed to be a result of local resonance. In digital circuitry, the flip flop was the digital construct of randomly accessible memory, and there is evidence of tuples of cells that remain activated for short periods of time after activation, supporting the idea of resonance.
Both these brain features are distinct from the mechanisms of an LSTM cell, or any of the cells that comprise the artificial networks that are currently in common use and represented in Java, R, Python, and C libraries or accelerated in Intel or NVidia products.
Plenty of money is going into developing chips that more accurately model the activation behavior and variable interconnectedness of neurons in artificial cells. Some of the advancements may present advantages over the current artificial networks types beyond just speed and use in distributed parallel architectures. They may exhibit fundamental changes to network behavior at the cellular and regional levels like brains that support the engineering of cognitive features into AI products. The nature and magnitude of the return on those investments is anyone's guess.