has been around for more than 60 years since the concept of AI was first proposed in 1956. Today, with the continuous innovation of related theories and technologies, AI is increasingly entering our daily life under the support of the "three elements" of data, computing power and algorithms.
However, behind this series of surprises, most AIs are struggling in terms of language understanding, visual scene understanding, decision analysis, etc.: these technologies are still mainly focused on the perception level, that is, using AI to simulate human hearing, vision and other perceptions However, it cannot solve complex cognitive and intelligent tasks such as reasoning, planning, association, and creation.
The current AI lacks the processing, understanding and thinking of information after entering the "brain". What it does is relatively simple comparison and identification, and only stays in the "perception" stage, not "cognition". The AI of the Lord is still far from human intelligence.
The reason is that AI is facing a bottleneck restricting its development: large-scale commonsense knowledge base and cognitive-based logical reasoning. The cognitive graph based on knowledge graph, cognitive reasoning, logic expression is considered by more and more domestic and foreign scholars and industry leaders as "one of the feasible solutions that can break through this technical bottleneck at present".
Recently, Gadi Singer, vice president of Intel Labs and named one of the 50 global thought leaders and influencers in the AI field, published an article titled The Rise of Cognitive AI, exploring artificial intelligence The third wave of : the rise of cognitive artificial intelligence.
Without changing the gist of the original text, Academic Toutiao has carefully compiled the article as follows:
Deep Learning (DL) is making huge strides and revolutionizing entire industries in every aspect of our lives, including healthcare , retail, manufacturing, autonomous vehicles, security and fraud prevention, and data analytics. However, in order to build the future of artificial intelligence (AI) and drive the next generation of technology further, we need to set a set of goals and expectations for it - by 2025, artificial intelligence will take a qualitative leap, and machines will be significantly changed. be smarter.
Currently, many applications based on deep learning algorithms address related perceptual tasks, such as object recognition, natural language processing (NLP), translation, and other tasks that involve extensive association of data (such as recommendation systems). in-depth learning system relies on differential programming and complex data-based correlation to make excellent results , and is expected to drive the transformation of the entire industry in the next few years. But at the same time, we must overcome the inherent limitations of deep learning itself to further help machine learning, or AI more broadly, to realize its potential. Achieving non-incremental innovation requires concerted efforts on three fronts:
- Substantially improve the efficiency of the model (e.g. reduce the number of parameters by 2-3 orders of magnitude without reducing its accuracy);
- Greatly enhances the robustness, scalability and scalability of the model;
- comprehensively improves the machine's cognitive ability .
has been around for more than 60 years since the concept of AI was first proposed in 1956. Today, with the continuous innovation of related theories and technologies, AI is increasingly entering our daily life under the support of the "three elements" of data, computing power and algorithms.
However, behind this series of surprises, most AIs are struggling in terms of language understanding, visual scene understanding, decision analysis, etc.: these technologies are still mainly focused on the perception level, that is, using AI to simulate human hearing, vision and other perceptions However, it cannot solve complex cognitive and intelligent tasks such as reasoning, planning, association, and creation.
The current AI lacks the processing, understanding and thinking of information after entering the "brain". What it does is relatively simple comparison and identification, and only stays in the "perception" stage, not "cognition". The AI of the Lord is still far from human intelligence.
The reason is that AI is facing a bottleneck restricting its development: large-scale commonsense knowledge base and cognitive-based logical reasoning. The cognitive graph based on knowledge graph, cognitive reasoning, logic expression is considered by more and more domestic and foreign scholars and industry leaders as "one of the feasible solutions that can break through this technical bottleneck at present".
Recently, Gadi Singer, vice president of Intel Labs and named one of the 50 global thought leaders and influencers in the AI field, published an article titled The Rise of Cognitive AI, exploring artificial intelligence The third wave of : the rise of cognitive artificial intelligence.
Without changing the gist of the original text, Academic Toutiao has carefully compiled the article as follows:
Deep Learning (DL) is making huge strides and revolutionizing entire industries in every aspect of our lives, including healthcare , retail, manufacturing, autonomous vehicles,Security and Fraud Prevention and Data Analytics. However, in order to build the future of artificial intelligence (AI) and drive the next generation of technology further, we need to set a set of goals and expectations for it - by 2025, artificial intelligence will take a qualitative leap, and machines will be significantly changed. be smarter.
Currently, many applications based on deep learning algorithms address related perceptual tasks, such as object recognition, natural language processing (NLP), translation, and other tasks that involve extensive association of data (such as recommendation systems). in-depth learning system relies on differential programming and complex data-based correlation to make excellent results , and is expected to drive the transformation of the entire industry in the next few years. But at the same time, we must overcome the inherent limitations of deep learning itself to further help machine learning, or AI more broadly, to realize its potential. Achieving non-incremental innovation requires concerted efforts on three fronts:
- Substantially improve the efficiency of the model (e.g. reduce the number of parameters by 2-3 orders of magnitude without reducing its accuracy);
- Greatly enhances the robustness, scalability and scalability of the model;
- comprehensively improves the machine's cognitive ability .
Figure | The number of parameters in deep learning-based language models increases exponentially (source: microsoft)
Although pruning, sparsity, compression, distillation and Figure neural network Techniques such as (GNN) can increase model efficiency, but ultimately yield incremental improvements at the same time. Reducing model size by orders of magnitude, without affecting results, may require more fundamental changes in the methods of capturing and representing the information itself and the ability to learn in deep learning models. also,Continued progress also requires more computationally efficient deep learning methods or a shift to other machine learning methods. Now, a promising class of artificial intelligence systems is rapidly gaining traction by retrieving in auxiliary information repositories to replace the embedding of large amounts of facts and data.
At the same time, statistical machine learning methods are based on the assumption of - that the distribution of training samples represents what has to be dealt with during inference, and has significant flaws in real-life use. Especially when the training dataset is sparsely sampled, or even lacks samples, the deep learning model will be challenged.
In addition, the results obtained with transfer learning and few-shot/zero-shot inference are not satisfactory. The inefficient scaling of models prevents AI from scaling to datasets and many areas that data scientists lack. In addition, deep learning is also very susceptible to data changes, resulting in low-confidence classifications, but this problem can be solved by improving the robustness and scalability of the model.
Finally, in most cases, neural networks fail to properly provide cognition, reasoning, and interpretability. Deep learning lacks cognitive mechanisms to reason about abstract , context context, causality, interpretability, and intelligibility.
The next stage: cognitive artificial intelligence
Artificial intelligence is expected to reach the level of human understanding. Relying on Daniel Kahneman's "Thinking, Fast and Slow" book to define the paradigm, Yoshua Bengio equates the capabilities of contemporary deep learning with what he describes as "System 1" features—intuitive, Rapid, unconscious, habitual and completely in a state of voluntary control. On the contrary, he pointed out,The next challenge for AI systems is to implement the functions of "System 2" - slow, logical, sequential, conscious and algorithmic, such as the functions required to implement planning and reasoning.
(Source: Pixabay)
Francois Chollet similarly describes an emerging stage in the development of artificial intelligence ("Flexible AI") on the basis of broad generalizations, which is able to adapt to unknown events in a wide range of fields. Both of these features are consistent with DARPA's (Defense Advanced Research Projects Agency) "third wave of artificial intelligence," which is characterized by contextual adaptation, abstraction, reasoning, and interpretability. One possible way to achieve these features is to combine deep learning with symbolic reasoning and deep knowledge . Below, I will use the term "Cognitive AI" to refer to this new phase of artificial intelligence.
Although we have no hope of open artificial general intelligence (AGI), artificial intelligence with higher cognitive capabilities can also play a greater role in technology and business. Once AI can make reliable decisions in unpredictable environments, it will eventually gain greater autonomy and play an important role in areas such as robotics technology, automated transportation, and control points in logistics, industrial and financial systems effect.
The role of structured knowledge in cognitive artificial intelligence
In the field of artificial intelligence, some believe that higher-level machine intelligence can be achieved by further developing deep learning, while others believe that this requires incorporating other fundamental mechanisms. In this regard, I agree with the latter for the following reasons:
Deep learning masters the statistical-based mapping from the input of the multidimensional structure in the embedding space to the predicted output. This allows it to distinguish between wide and shallow data (for example,good performance in terms of words or pixel/voxel sequences in images). Furthermore, deep learning is equally effective at indexing resources (like Wikipedia) and retrieving answers from the best match in the corpus - as demonstrated on benchmarks such as NaturalQA or EffiicentQA. As defined by Bengio, the task of System 1 relies on statistical mapping functions created during training. And deep learning can help with these tasks.
By contrast, structured, explicit, and understandable knowledge can provide a path to more advanced machine intelligence or system 2 capabilities. A fundamental knowledge building is the ability to capture declarative knowledge about elements and concepts and encode abstract concepts (eg, hierarchical attribute inheritance between classes). For example, knowledge about birds, plus information about passerines, plus detailed information about sparrows, provides a lot of implicit information about chestnut sparrows, even if not specifically stated. In addition to this, other knowledge constructs include causal and predictive models. Constructions such as
rely on explicit concepts and well-defined relationships rather than embedded machines in the latent space, and the resulting models will therefore have a wider range of explanatory and predictive potential beyond the capabilities of statistical mapping.
(Source: Pixabay)
The human brain has the ability to "imagine," simulate, and evaluate potential future events that are beyond the reach of experience or observation. At the same time, these capabilities provide an evolutionary advantage to human intelligence. In an environment unconstrained by explicit rules, mental simulations of possible future events are based on fundamental models of world dynamics, which have great adaptive value in planning and problem solving.
Process modeling mechanisms are based on implicit mathematical, physical or psychological principles rather than observable statistical correlations from input to output,This is critical for achieving higher cognitive abilities. For example, physical models can capture hydroplaning phenomena and make simple predictions about the motion of a car under various conditions. Such process models can be used in conjunction with deep learning-based approaches to extend the capabilities of current AI. The
knowledge base can capture (or implicitly) common-sense assumptions and underlying logic that are not always publicly present in the training data of deep learning systems. This suggests that an understanding of the world and its dynamics can help solve tasks for higher-level machine intelligence. Finally, reasonably structured knowledge can be disambiguated in terms of contextual context and aggregated content (classifying the attributes of a "club" as baseball, weapons, cards, or meeting places).
Cognitive Artificial Intelligence and the Age of Knowledge
In the next few years, as shallow mapping functions become richer and computational processing becomes more economical and faster, deep learning-based systems 1 are expected to make significant progress. Cognitive AI will also bring more and more advanced capabilities.
All in all, I believe that by 2025, there will be a new crop of cognitive AIs that are not only more explanatory, but also closer to the level of human autonomous reasoning than current deep learning-based systems.
We have established a Cognitive Computing Research Unit at Intel Labs to drive Intel's innovation at the intersection of machine intelligence and cognition, and to continuously advance the capabilities of emerging cognitive artificial intelligence. We strive to combine the latest achievements in deep learning with the integration of knowledge building and neurosymbolic artificial intelligence to build self-learning artificial intelligence capable of making informed decisions in complex scenarios.
Deep learning enables artificial intelligence systems to excel in recognition, perception, translation and recommender system tasks. The rise of the next wave of machine learning and artificial intelligence technologies,A new type of artificial intelligence with greater understanding and cognition will be created to bring greater convenience to our lives.
The next stage: cognitive artificial intelligence
Artificial intelligence is expected to reach the level of human understanding. Relying on Daniel Kahneman's "Thinking, Fast and Slow" book to define the paradigm, Yoshua Bengio equates the capabilities of contemporary deep learning with what he describes as "System 1" features—intuitive, Rapid, unconscious, habitual and completely in a state of voluntary control. In contrast, he noted, the next challenge for AI systems lies in implementing "System 2" functions -- slow, logical, sequential, conscious and algorithmic, such as those required to implement planning and reasoning.
(Source: Pixabay)
Francois Chollet similarly describes an emerging stage in the development of artificial intelligence ("Flexible AI") on the basis of broad generalizations, which is able to adapt to unknown events in a wide range of fields. Both of these features are consistent with DARPA's (Defense Advanced Research Projects Agency) "third wave of artificial intelligence," which is characterized by contextual adaptation, abstraction, reasoning, and interpretability. One possible way to achieve these features is to combine deep learning with symbolic reasoning and deep knowledge . Below, I will use the term "Cognitive AI" to refer to this new phase of artificial intelligence.
Although we have no hope of open artificial general intelligence (AGI), artificial intelligence with higher cognitive capabilities can also play a greater role in technology and business. Once AI can make reliable decisions in unpredictable environments, it will eventually gain greater autonomy and play an important role in areas such as robotics technology, automated transportation, and control points in logistics, industrial and financial systems effect.
The role of structured knowledge in cognitive artificial intelligence
In the field of artificial intelligence, some believe that higher-level machine intelligence can be achieved by further developing deep learning, while others believe that this requires incorporating other fundamental mechanisms. In this regard, I agree with the latter for the following reasons:
Deep learning masters the statistical-based mapping from the input of the multidimensional structure in the embedding space to the predicted output. This makes it good at distinguishing between wide and shallow data (e.g. words or pixel/voxel sequences in images). Furthermore, deep learning is equally effective at indexing resources (like Wikipedia) and retrieving answers from the best match in the corpus - as demonstrated on benchmarks such as NaturalQA or EffiicentQA. As defined by Bengio, the task of System 1 relies on statistical mapping functions created during training. And deep learning can help with these tasks.
By contrast, structured, explicit, and understandable knowledge can provide a path to more advanced machine intelligence or system 2 capabilities. A fundamental knowledge building is the ability to capture declarative knowledge about elements and concepts and encode abstract concepts (eg, hierarchical attribute inheritance between classes). For example, knowledge about birds, plus information about passerines, plus detailed information about sparrows, provides a lot of implicit information about chestnut sparrows, even if not specifically stated. In addition to this, other knowledge constructs include causal and predictive models. Constructions such as
rely on explicit concepts and well-defined relationships rather than embedded machines in the latent space, and the resulting models will therefore have a wider range of explanatory and predictive potential beyond the capabilities of statistical mapping.
(Source: Pixabay)
The human brain has the ability to "imagine," simulate, and evaluate potential future events that are beyond the reach of experience or observation. At the same time, these capabilities provide an evolutionary advantage to human intelligence. In an environment unconstrained by explicit rules, mental simulations of possible future events are based on fundamental models of world dynamics, which have great adaptive value in planning and problem solving.
Process modeling mechanisms are based on implicit mathematical, physical, or psychological principles, rather than observable statistical correlations from input to output, which are critical for achieving higher cognitive abilities. For example, physical models can capture hydroplaning phenomena and make simple predictions about the motion of a car under various conditions. Such process models can be used in conjunction with deep learning-based approaches to extend the capabilities of current AI. The
knowledge base can capture (or implicitly) common-sense assumptions and underlying logic that are not always publicly present in the training data of deep learning systems. This suggests that an understanding of the world and its dynamics can help solve tasks for higher-level machine intelligence. Finally, reasonably structured knowledge can be disambiguated in terms of contextual context and aggregated content (classifying the attributes of a "club" as baseball, weapons, cards, or meeting places).
Cognitive Artificial Intelligence and the Age of Knowledge
In the next few years, as shallow mapping functions become richer and computational processing becomes more economical and faster, deep learning-based systems 1 are expected to make significant progress. Cognitive AI will also bring more and more advanced capabilities.
All in all, I believe that by 2025, there will be a new crop of cognitive AIs that are not only more explanatory, but also closer to the level of human autonomous reasoning than current deep learning-based systems.
We have established a Cognitive Computing Research Unit at Intel Labs to drive Intel's innovation at the intersection of machine intelligence and cognition, and to continuously advance the capabilities of emerging cognitive artificial intelligence. We strive to combine the latest achievements in deep learning with the integration of knowledge building and neurosymbolic artificial intelligence to build self-learning artificial intelligence capable of making informed decisions in complex scenarios.
Deep learning enables artificial intelligence systems to excel in recognition, perception, translation and recommender system tasks. The rise of the next wave of machine learning and artificial intelligence technology will create a new type of artificial intelligence with greater understanding and cognition, which will bring greater convenience to our lives.
