8 AI Technologies Beyond Neural Networks

Big Data Digest Works

Compiled by: Fu Yiyang, Ding Hui, Aileen

Currently, the most funding in the AI technology field is being directed towards research on neural networks. To many, neural network technology seems to be the “brain constructed by programs” (though this analogy is quite inaccurate).

The concept of neural networks was proposed as early as the 1940s, but even now, we know very little about how neurons and the brain work. In recent years, there has been a growing call in the scientific community for innovation in neural network technology, eager to reignite the neural network wave…

In fact, besides neural networks, the AI field includes many more interesting, novel, and promising technologies, which this article will introduce to everyone.

Knol refers to information units, such as keywords and phrases. The knol extraction technology is the process of extracting key information from text. For example, the sentence “As the name suggests, an octopus has 8 legs” would be transformed into: {“octopus”: {“number of legs”: 8}}.

The Google search engine we commonly use relies on this technology, and many of the technologies introduced later also include this technology.

Ontology construction is an NLP-based technology aimed at building a hierarchy of entity nouns with software, which greatly aids in achieving AI conversations. Although ontology construction seems simple on the surface, it is actually quite complex, mainly because the actual relationships between things are much more complicated than we think.

For example, using NLP to analyze text to establish entity relationship sets:

Example sentence: “My Labrador just had a litter of puppies, and their father is a lion dog, so they are Labrador Poodles (a mixed breed)”. This sentence would be transformed into: {“puppies”: {“may be”: “Labrador Poodle”, “have”: “father”}, “Labrador”: {“have”: “puppies”}}.

However, when humans express language, they often do not state all relationships. For example, in this sentence, one must infer that “my Labrador is female”. This is the difficulty of ontology construction.

Thus, ontology construction technology is currently only applied in top-tier chatbots.

A heuristic is a rule used for classification, often resembling conditional statements like “if this item is red” or “if Bob is at home”. These conditional statements are often accompanied by an action or decision, for example:

If an item’s [“ingredient”] attribute contains the element “arsenic”:

Then its [“poison”] attribute is “True”.

For each new piece of information, new heuristics and new relationships are introduced. As new heuristics are established, new understandings of related nouns can emerge. For instance:

Heuristic 1: “puppies” indicate they are young;

Heuristic 2: Young indicates they are very young;

From these two heuristics, one can infer that “puppies” are all very young.

The difficulty with heuristics is that in most cases, the rules are not as simple as “If/Then”. Statements like “some people have blonde hair” are hard to express with heuristics. Hence, we have “epistemology” (see below).

Epistemology combines ontology construction and custom heuristics, incorporating probabilistic characteristics to represent the likelihood of a noun being associated with any attribute.For example, using the following ontology structure:

{‘person’: {‘gender’: {‘male’: 0.49, ‘female’: 0.51}, ‘race’: {‘Asian’: 0.6, ‘African’: 0.14}}

to represent the judgment of a person’s gender and race. Additionally, probabilities can help identify some “mixed” phrases with multiple meanings, for instance, in the sentence “The plum looks like a hormone-injected raisin”, the phrase “hormone-injected” likely means “larger size”, leading to the conclusion that the sentence likely means “the plum is larger than the raisin”.

Implementing epistemology is much more challenging than ontology construction. First, it requires more data; and due to its structural complexity, it is difficult to quickly establish a database for lookup after determining the rules. Moreover, the determination of rules is usually based on the frequency with which a thing is mentioned in a piece of text, but the text may not accurately reflect reality.

Epistemology is quite similar to the “tensor flow” theory proposed by Asimov. The TensorFlow system developed by Google is not truly based on tensors, while epistemology is based on tensors.

A metric system must include corresponding evaluation standards. Imagine when purchasing a house, factors such as area, location, price, and style need to be considered, and these factors may not all be positive, requiring a decision through weighing trade-offs. For instance, if you care more about area than price, you might be willing to spend several times more to buy a larger house.

Self-assessment technology determines the weight of each factor based on your importance level for different factors, thereby providing decision suggestions.Through this process, it can also predict inventory changes, recommend products, and achieve autonomous driving, meaning that most functions achievable by neural networks can also be performed by automated metric technology, albeit requiring longer training times, but with decision speeds several orders of magnitude faster.

Vector differencing technology is commonly used in image analysis and can also be applied to processing time-varying data. By constructing an abstract vector graph of the target, candidate objects are compared with the target object to determine if it is the “best face shape for a date” or the “best time to buy”.

Typically, differences between target objects come with a quantification rule to measure the degree of difference, allowing some “fuzzy” concepts to be represented simply and clearly through vectorization of features.

For example, for humans, we generally believe that symmetrical faces are more attractive, but for computers, precise calculations are needed to make judgments. In this case, abstracting the face using 30 triangles saves much computation time and storage space compared to using the full facial image for comparison.

Non-image data processing is also possible. For example, changes in stock prices, earnings per share, and margin ratios can be vectorized and compared with ideal values to determine the benefits or risks of an investment.

Convolution matrices are often used in image processing for edge detection and contrast enhancement, for example, many filters in Photoshop are based on convolution matrices or stacked convolutions (multiple convolution operations performed in a specific order).

Additionally, convolution matrices can be used to process non-image data. For example, when processing time series vectors with convolution matrices, patterns can be quickly identified like edge detection, and specific values or ranges can be found at minimum or maximum points to make judgments.

Making a decision is not simple. A multi-perspective decision system makes decisions in a more democratic way, considering multiple aspects.

For instance, in the house example, your positive outlook on a particular house might be based on incomplete factors, and later a fact like “this house is built on a cliff” (of course, this overwhelming factor might come from knol extraction) would erase all your prior favor, prompting you to reconsider your decision.

Therefore, decisions need to be made considering a more comprehensive set of factors, and a multi-perspective decision system can utilize two sets of standards from two individuals (like you and your spouse) to evaluate decisions. Multi-perspective decision systems can also be applied in autonomous driving, for example, collecting the opinions of 10,000 car owners to establish new standards.

Many people see only one tool, falling into the trap of “I have a hammer, so everything is a nail.” Companies like Recognant apply neural networks while also utilizing these relatively obscure technologies mentioned in the article because compared to neural network hardware systems,

the advantage of these software technologies lies in their ability to adjust and develop for different situations at any time without incurring additional costs.Thus, having a narrow technical scope can lead to being trapped by certain situations, while a broader technical scope makes it easier to tackle problems.

Original link:

https://www.linkedin.com/pulse/8-ai-technologies-aint-neural-networks-brandon-wirtz/

[Today’s Machine Learning Concept]

Have a Great Definition