VICARIOUS

Vicarious is a AI company created in 2010 by Scott Phoenix and Dileep George and located in San Francisco Bay Area, California. This company was acquired in 2022 by Alphabet Inc.

It is based on recursive cognition theory which is the idea that the brain uses a process of recursive reasoning to understand the world. The aim of the company is to create robots that can understand and generate natural language, and perform tasks like object manipulation and image recognition. Vicarious is on the path to creating robots that have the ability to think, behave, and learn like humans.

Yummyvite Green Tea Extract is developed by PureMaxLabs. It contains natural caffeine and L-theanine contained in green tea and it provides energy and improves brain function. This supplement is also packed with polyphenol catechins and epigallocatechin which promote healthy cells and immune systems. Buy here.

Recursive Cognition

This theory is based on the human brain and how it processes information. It reveals that the brain utilizes a process of recursive reasoning to comprehend what happens around us. It is based on the premise that the brain exists as a hierarchical system, and each level of the hierarchy represents a different level of complexity or abstraction. That is, mental representations can be seen as building blocks combined in several ways to develop a more complex representation.  Hence, the brain uses recursive cognition by breaking down complex information into simpler components and re-combining those components to create a complete understanding. For example, the brain can take a simple representation of an object and recursively combine it with other representations to develop a more convoluted representation of the object’s properties in relation to other objects.

With recursive cognition, the brain creates a model of the world which is used to make predictions, solve problems, make decisions, and other cognitive activities that enable humans to understand natural language and perform tasks that require human intelligence. For example, the mind may create a hypothesis and use recursive processes to test and enhance the hypothesis to find a solution. In addition, the mind also uses this process to develop categories and concepts through the combination of different object features to create a general concept.

Over the years, researchers have been trying to develop computational models of recursive cognition to imitate the recursive operations of the mind. For example, the development of artificial neural networks in AI. This is what Vicarious technology company is based on because the company tries to replicate the process of recursive cognition theory in AI systems.

Recursive Cortical Network (RCN)

Vicarious aims at creating an algorithm that will be able to detect certain letters always and in every setting. Detecting letters under several situations can be difficult for machines. Due to this, CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) was formed to differentiate between humans and robots. On October 22, 2013, Vicarious announced that its model was reliably able to crack CAPTCHA using deep neural networks. Over the years, they have improved, and their software now requires fewer data to generalize all CAPTCHAs.

In other to achieve this, Vicarious developed machine learning (ML) software that depends on the human brain’s computational principles. This software is a generative graphical visual perception system called Recursive Cortical Network which interprets photographs and videos’ contents in a way similar to humans. RCN constructs a data-generating model or generative model that stipulates how an image is derived from a group of latent variables. The data generating model can be used in data sampling similar to the data it is modeling or used to deduce latent variables, that is, influence by generation. This means that RCN is not a neural network but a probabilistic graphical model. For example, it separates textures of objects from their shape and objects from backgrounds.

What makes RCN different from a normal neural network is that a normal neural network starts without any knowledge before training, while, RCN starts with knowledge of contours and surfaces. This knowledge facilitates generalizability and model building. For example, where a normal neural network would have to learn what contours and surfaces exist and has to know the difference that exists between objects and backgrounds, an RCN already knows. This makes it easier to train an RCN on many individual objects.

The development of RCN comes from cognitive science and neuroscience. Humans with neocortex at birth can already differentiate between foreground and background. This facilitates learning the representations in our world compared to if we had to start learning everything from a clean slate.

An RCN is modeled as a mix of surfaces and contours. The former are models which use a conditional random field while the latter uses a compositional hierarchy of features. This allows RCN to be able to recognize shapes of objects with different appearances without training on all the combinations of shapes and appearances. In addition, a good feature of RCNs is that they do not over-fit to extraneous details of the training set. That is, RCNs have a strong generalization compared to other similar scenes.

The generalization of ML techniques is crucial in solving CAPTCHAs. Vicarious has realized that a slight disturbance in the CAPTCHA will render their traditional neural network approach useless.

Yummyvite Green Tea Extract is developed by PureMaxLabs. It contains natural caffeine and L-theanine contained in green tea and it provides energy and improves brain function. This supplement is also packed with polyphenol catechins and epigallocatechin which promote healthy cells and immune systems. Buy here.

Learning an RCN

Pre-processing

This stage is important when learning an RCN. In this stage, the training image passes through an edge detector to separate the edges. To achieve this, Gabor filters are used. These filters are oriented and are activated under the presence of a specific edge at a precise orientation. The number of filters is normally fixed at 16. However, if you need more intricacy, it can be fixed to a higher number. The output gotten when the image passes through the filters should be in an array in the form of (num filters, width, height), with each (width, height) element of the array containing edges of that specific edge orientation.

Sparsification

RCNs do not learn through backpropagation because they learn through dictionary learning. The dictionary learning algorithm greedily sparsifies the edge map through the detection of edge activations and suppressing other activations in the radius. The detected edge activations are stored in a dictionary as (f, r, c) tuples, which are in line with the feature index (edge orientation), row, and column of the activated edge consecutively.

Lateral Learning

Lateral learning is crucial in the operation of how the visual cortex operates. After learning the latent variables through dictionary learning, a graph that constrains the model is constructed. Lateral learning is required because the graphs do not add any randomness to the model, hence, images sampled from the model will not have any variation. This is done through a pooling layer which arbitrarily moves the activated features inside a specific radius. If there is no constrainment, the resulting image may have excessive variation and may not look like the image it is to model. This is prevented from happening with lateral connections. Lateral connections are created through the addition of pairwise edges between features from the closest to the longest. The resulting graph consists of both short and long connections, and it should be identical to the image being modeled.

Inference in the RCN

After the learning stage is completed, you can move to the next stage which is the influence stage. Whenever you are carrying out the influence stage, you are not just limited to classification but generation and segmentation. Generation can be done because it is a generative model while segmentation is through top-down attention which can be down through explaining away and lateral connections. This makes it suitable to be used for applications that need a flexible vision system, such as robots.

Inference involves a two-step procedure;

Forward Pass

RCNs carry out Bayesian inference which is time-consuming and expensive. On loopy graphs, it becomes more difficult because it is impossible to search for the exact marginal. This requires approximations to be made instead. However, message-passing belief propagation enables forward passes, and an adequate approximation can be located by using some neat tricks.

When using the forward pass, the input image goes through the Gabor filters. Then, a graph cut is done to convert the loopy graph into a tree. This makes it possible for the precise inference to be carried out quickly through every graph using the max-product algorithm. Generally, the forward pass is all that is required, but it can overestimate the marginal and make wrong assumptions. Hence, the forward pass can be used to select a group of high-scoring graphs or a set of candidate hypotheses.

Backward Pass

Here, candidates selected in the forward pass will be refined using loopy belief propagation on the full graph. Also, several multiple iterations of loopy belief propagation are carried out to explain any conflicting variables. When this is completed, the backtraced latent variables (f, r, c) will be decoded into the edge map, leading to top-down attention.

Benefits of Vicarious Technology

Enhanced Automation; with vicarious technology, AI systems will be able to carry out several tasks autonomously which will enhance efficiency and productivity in all industries such as manufacturing, logistics, and healthcare.

Cost Reduction; labor costs will be reduced and this will cut costs for companies.

Enhanced Accuracy; AI systems are trained using massive amounts of data which improves their performance and reliability over time.

Safety; robotic systems can work under hazardous conditions which will increase safety for humans.

Flexibility; AI systems are not rigid as they are more versatile and can adapt to new conditions, unlike traditional robotic systems.

Cons of Vicarious Technology

Ethical Concerns; vicarious technology is not really ethical due to the numerous issues with autonomy and privacy.

Understanding the Human Brain; the present understanding of the human brain is still very much limited, and this poses a problem because these limitations make it difficult to create reliable and accurate models.

High Computational Requirements; vicarious technology uses a large amount of data and computational power, and this makes it expensive and resource intensive.

Lack of Transparency; due to the complexity and nonlinearity of the models, it can be difficult to understand and interpret the behaviour of a vicarious model, and this can make it hard to trust the decisions taken by the model.

Limited Scalability; the complexity of the computational requirements needed by vicarious technology makes it difficult to scale up the vicarious models to manage large-scale data sets or tasks.

Bad Intentions: people with bad intentions might end up misusing this technology for malicious purposes such as manipulation, surveillance, and control.

Limited Applicability; vicarious technology can only mimic a few biological systems.

Yummyvite Green Tea Extract is developed by PureMaxLabs. It contains natural caffeine and L-theanine contained in green tea and it provides energy and improves brain function. This supplement is also packed with polyphenol catechins and epigallocatechin which promote healthy cells and immune systems. Buy here.

Conclusion

Vicarious technology still has a long way to go. It is promising, and still under development, so it is important to carry out your research and understand the possible benefits, challenges, and limitations before implementation. Additionally, understand the broader consequences of AI-controlled robots and their effects on people and society as a whole.

Important Affiliate Disclosure

We at culturedlink.com are esteemed to be a major affiliate for some of these products. Therefore, if you click any of these product links to buy a subscription, we earn a commission. However, you do not pay a higher amount for this. Rest easy as the information provided here is accurate and dependable.